AI Story Market: Evolving Writer Agents for Short-Drama IP

Inspiration

Short-form video content is exploding on TikTok, YouTube Shorts, and Reels, but creating compelling narratives at scale is expensive and time-consuming. We asked: What if AI agents could compete, evolve, and improve their storytelling abilities through reader feedback?

We were inspired by:

The rise of short drama content in Asian markets (billions of views)
Evolutionary algorithms and competitive learning systems
The need for scalable IP creation for video production

What it does

SHORT DRAMA IP FACTORY: a competitive AI storytelling platform where multiple writer agents battle to create the most engaging short drama scripts.

How we built it

🎭 Competitive Writing System

5+ unique writer agents with distinct storytelling styles
Reader/critic agents evaluate on Novelty, Alignment, Coherence
Multi-dimensional scoring system with composite ratings

📊 Evolution Visualization

Interactive charts tracking writer performance across rounds
See which agents improve, plateau, or fail over time
Click any data point for deep story analysis

📖 Interactive Story Viewer

Read full scripts with professional formatting
View detailed reader feedback with score breakdowns
Side-by-side story comparison

🎬 AI Video Export Pipeline

OpenAI GPT-4o generates shot-by-shot beat sheets
Seedance (ByteDance) AI for real video generation
Export for Runway, Pika, Sora video platforms

🔐 Production Features

Privy wallet authentication for future prediction markets
Sentry error tracking across frontend/backend
Deployed on Vercel + Render

Specifically Frontend Stack

Next.js 16 with React 19 - Latest performance features
TypeScript - Full type safety
Tailwind CSS 4 - Brutalist design aesthetic
Recharts - Interactive data visualization
Privy - Embedded Solana wallet authentication

Backend Stack

FastAPI - High-performance Python framework
OpenAI GPT-4o - Beat sheet generation & writer agents
fal.ai SDK - Seedance video generation
Pydantic - Data validation
Sentry - Production monitoring

Deployment

Vercel for frontend (auto-deploy from main)
Render for backend (environment management)
GitHub PR-based workflow

Architecture We built a two-tier system: Next.js frontend visualizing agent evolution + FastAPI backend running competitive writing rounds with real AI integration (OpenAI, Seedance).

Challenges we ran into

Writer Agent Challenges: LLMs tend to mode-collapse into the same kind of story, same distribution, even same character names. LLMs when tasked to do two things at once, will get confused and only accomplish one. iteration speed gets slower as we add more complicated tasks.

Reader Agent Challenges: AI being overly positive AI opinions not diverse enough Conflict between meeting multiple objectives: e.g., story needs to resolve established conflict but also not falling into audience's expectation

Full Stack Challenges: Seedance Video Integration - Version conflicts (fal-client 0.6.2 didn't exist), API parameter mapping struggles, deployment timing issues. Eventually got real video generation working in production.

Accomplishments that we're proud of

Writer Agent Accomplishments: effective context management systems good coding practices adaptable to new ideas.

Reader Agent Accomplishments: An expectation-based novelty evaluation mechanism AI audience producing pretty good feedback

Fullstack accomplishments

AI Integration - OpenAI GPT-4o for beat sheets, custom multi-agent competitive system
Beautiful, Accessible UI - Brutalist design with interactive charts, comprehensive reader feedback, smooth animations
Production Infrastructure - Sentry error tracking, Privy wallet auth, deployed on Vercel + Render with proper CORS